{"id":16388205,"url":"https://github.com/kentonishi/augmentation-for-lnl","last_synced_at":"2025-03-16T16:31:08.091Z","repository":{"id":41472029,"uuid":"313829882","full_name":"KentoNishi/Augmentation-for-LNL","owner":"KentoNishi","description":"[CVPR 2021] Code for \"Augmentation Strategies for Learning with Noisy Labels\".","archived":false,"fork":false,"pushed_at":"2022-01-09T06:16:57.000Z","size":52500,"stargazers_count":112,"open_issues_count":0,"forks_count":13,"subscribers_count":6,"default_branch":"master","last_synced_at":"2025-02-27T11:36:31.715Z","etag":null,"topics":["augmentation-policies","cifar10","cifar100","clothing1m","cvpr","cvpr2021","data-augmentation","data-augmentation-strategies","label-noise","label-noise-robustness","semi-supervised-learning"],"latest_commit_sha":null,"homepage":"https://openaccess.thecvf.com/content/CVPR2021/html/Nishi_Augmentation_Strategies_for_Learning_With_Noisy_Labels_CVPR_2021_paper.html","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"mit","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/KentoNishi.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null}},"created_at":"2020-11-18T05:12:40.000Z","updated_at":"2025-02-27T07:30:12.000Z","dependencies_parsed_at":"2022-09-12T00:50:57.210Z","dependency_job_id":null,"html_url":"https://github.com/KentoNishi/Augmentation-for-LNL","commit_stats":null,"previous_names":[],"tags_count":0,"template":false,"template_full_name":null,"repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/KentoNishi%2FAugmentation-for-LNL","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/KentoNishi%2FAugmentation-for-LNL/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/KentoNishi%2FAugmentation-for-LNL/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/KentoNishi%2FAugmentation-for-LNL/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/KentoNishi","download_url":"https://codeload.github.com/KentoNishi/Augmentation-for-LNL/tar.gz/refs/heads/master","host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":243822283,"owners_count":20353499,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2022-07-04T15:15:14.044Z","host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["augmentation-policies","cifar10","cifar100","clothing1m","cvpr","cvpr2021","data-augmentation","data-augmentation-strategies","label-noise","label-noise-robustness","semi-supervised-learning"],"created_at":"2024-10-11T04:28:35.496Z","updated_at":"2025-03-16T16:31:06.543Z","avatar_url":"https://github.com/KentoNishi.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Augmentation-for-LNL\n\n[![PWC](https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/augmentation-strategies-for-learning-with/image-classification-on-clothing1m)](https://paperswithcode.com/sota/image-classification-on-clothing1m?p=augmentation-strategies-for-learning-with)\n\nCode for ***Augmentation Strategies for Learning with Noisy Labels*** (CVPR 2021).\n\n**[Authors](mailto:kento24gs@outlook.com,yding@cs.ucsb.edu,anrich@cs.ucsb.edu,holl@cs.ucsb.edu)**: [Kento Nishi](mailto:kento24gs@outlook.com)\\*, [Yi Ding](mailto:yding@cs.ucsb.edu)\\*, [Alex Rich](mailto:anrich@cs.ucsb.edu), [Tobias Höllerer](mailto:holl@cs.ucsb.edu) [`*`: [equal contribution](mailto:kento24gs@outlook.com,yding@cs.ucsb.edu)]\n\n\u003cdetails\u003e\n    \u003csummary\u003eAbstract\u003c/summary\u003e\n    Imperfect labels are ubiquitous in real-world datasets. Several recent successful methods for training deep neural networks (DNNs) robust to label noise have used two primary techniques: filtering samples based on loss during a warm-up phase to curate an initial set of cleanly labeled samples, and using the output of a network as a pseudo-label for subsequent loss calculations. \n    In this paper, we evaluate different augmentation strategies for algorithms tackling the \"learning with noisy labels\" problem. We propose and examine multiple augmentation strategies and evaluate them using synthetic datasets based on CIFAR-10 and CIFAR-100, as well as on the real-world dataset Clothing1M. \n    Due to several commonalities in these algorithms, we find that using one set of augmentations for loss modeling tasks and another set for learning is the most effective, improving results on the state-of-the-art and other previous methods. Furthermore, we find that applying augmentation during the warm-up period can negatively impact the loss convergence behavior of correctly versus incorrectly labeled samples. We introduce this augmentation strategy to the state-of-the-art technique and demonstrate that we can improve performance across all evaluated noise levels. In particular, we improve accuracy on the CIFAR-10 benchmark at 90% symmetric noise by more than 15% in absolute accuracy, and we also improve performance on the real-world dataset Clothing1M.\n\u003c/details\u003e\n\n\u003cp align=\"center\"\u003e\n    \u003cimg src=\"./banner.png\" alt=\"Banner\" /\u003e\n\u003c/p\u003e\n\n[View on arXiv](https://arxiv.org/abs/2103.02130) / [View PDF](https://arxiv.org/pdf/2103.02130.pdf) / [Download Paper Source](https://arxiv.org/e-print/2103.02130) / [Download Source Code](https://github.com/KentoNishi/Augmentation-for-LNL/archive/master.zip)\n\n\u003cp align=\"center\"\u003e\n    \u003ca href=\"https://kentonishi.github.io/Augmentation-for-LNL/CVPR_Video.mp4\"\u003e\n        \u003cimg src=\"./thumbnail.png\" alt=\"Thumbnail\"/\u003e\n        \u003cbr /\u003e\n        Watch CVPR Video\n    \u003c/a\u003e\n\u003c/p\u003e\n\n## Benchmarks\n\u003cdetails\u003e\n    \u003csummary\u003eAll Benchmarks\u003c/summary\u003e\n    \u003ch3\u003eKey\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eAnnotation\u003c/th\u003e\n                \u003cth\u003eMeaning\u003c/th\u003e\n            \u003c/tr\u003e\n        \u003c/thead\u003e\n        \u003ctbody\u003e\n            \u003ctr\u003e\n                \u003ctd\u003e\u003ccode\u003eSmall\u003c/code\u003e\u003c/td\u003e\n                \u003ctd\u003eWorse or equivalent to previous state-of-the-art\u003c/td\u003e\n            \u003c/tr\u003e\n            \u003ctr\u003e\n                \u003ctd\u003eNormal\u003c/td\u003e\n                \u003ctd\u003eBetter than previous state-of-the-art\u003c/td\u003e\n            \u003c/tr\u003e\n            \u003ctr\u003e\n                \u003ctd\u003e\u003cstrong\u003eBold\u003c/strong\u003e\u003c/td\u003e\n                \u003ctd\u003eBest in task/category\u003c/td\u003e\n            \u003c/tr\u003e\n        \u003c/tbody\u003e\n    \u003c/table\u003e\n    \u003ch3\u003eCIFAR-10\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eModel\u003c/th\u003e\n                \u003cth\u003eMetric\u003c/th\u003e\n                \u003cth colspan=\"5\"\u003eNoise Type/Ratio\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctr\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e20% sym\u003c/th\u003e\n                \u003cth\u003e50% sym\u003c/th\u003e\n                \u003cth\u003e80% sym\u003c/th\u003e\n                \u003cth\u003e90% sym\u003c/th\u003e\n                \u003cth\u003e40% asym\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctbody\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eRuntime-W (Vanilla DivideMix)\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e96.100%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e94.600%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e93.200%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e76.000%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e93.400%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e95.700%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e94.400%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e92.900%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e75.400%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e92.100%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eRaw\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e85.940%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e27.580%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e83.230%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e23.915%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eExpansion.Weak\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e90.860%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e31.220%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e89.948%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e10.000%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eExpansion.Strong\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e90.560%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e35.100%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e89.514%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e34.228%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WW\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e96.270%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e36.050%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e96.084%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e23.503%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eRuntime-S\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e96.540%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e70.470%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e96.327%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e70.223%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-SS\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e96.470%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e81.770%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e96.193%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e81.540%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WS.RandAug.n1m6\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e96.280%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e89.750%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e96.006%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e89.629%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WS.SAW\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e96.350%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e95.640%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e93.720%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e35.330%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e94.390%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e96.138%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e95.417%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e93.563%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e10.000%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e94.078%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WS (WAW)\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e96.330%\u003c/td\u003e\n                    \u003ctd\u003e95.360%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e93.770%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e91.880%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e94.640%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e96.168%\u003c/td\u003e\n                    \u003ctd\u003e95.134%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e93.641%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e91.760%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e94.258%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n            \u003c/tbody\u003e\n        \u003c/thead\u003e\n    \u003c/table\u003e\n    \u003ch3\u003eCIFAR-100\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eModel\u003c/th\u003e\n                \u003cth\u003eMetric\u003c/th\u003e\n                \u003cth colspan=\"4\"\u003eNoise Type/Ratio\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctr\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e20% sym\u003c/th\u003e\n                \u003cth\u003e50% sym\u003c/th\u003e\n                \u003cth\u003e80% sym\u003c/th\u003e\n                \u003cth\u003e90% sym\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctbody\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eRuntime-W (Vanilla DivideMix)\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e77.300%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e74.600%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e60.200%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e31.500%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e76.900%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e74.200%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e59.600%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e31.000%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eRaw\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e52.240%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e7.990%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e39.176%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e2.979%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eExpansion.Weak\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e57.110%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e7.300%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e53.288%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e2.223%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eExpansion.Strong\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e55.150%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e7.540%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e54.369%\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e3.242%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WW\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e78.900%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e30.330%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e78.437%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e29.876%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eRuntime-S\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e79.890%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e40.520%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e79.395%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e40.343%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-SS\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e79.790%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e38.850%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e79.511%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e38.553%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WS.RandAug.n1m6\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e78.060%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e36.890%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e77.826%\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e\u003c/code\u003e\u003c/td\u003e\n                    \u003ctd\u003e36.672%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WS.SAW\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e79.610%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e77.640%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e61.830%\u003c/td\u003e\n                    \u003ctd\u003e17.570%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e79.464%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e77.522%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e61.632%\u003c/td\u003e\n                    \u003ctd\u003e15.050%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eAugDesc-WS (WAW)\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e79.500%\u003c/td\u003e\n                    \u003ctd\u003e77.240%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e66.360%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e41.200%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e79.216%\u003c/td\u003e\n                    \u003ctd\u003e77.010%\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e66.046%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e40.895%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n            \u003c/tbody\u003e\n        \u003c/thead\u003e\n    \u003c/table\u003e\n    \u003ch3\u003eClothing1M\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eModel\u003c/th\u003e\n                \u003cth\u003eAccuracy\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctbody\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eRuntime-W (Vanilla DivideMix)\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e74.760%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eAugDesc-WS (WAW)\u003c/td\u003e\n                    \u003ctd\u003e\u003ccode\u003e74.720%\u003c/code\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eAugDesc-WS.SAW\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e75.109%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n            \u003c/tbody\u003e\n        \u003c/thead\u003e\n    \u003c/table\u003e\n\u003c/details\u003e\n\n\u003cdetails open\u003e\n    \u003csummary\u003eSummary Metrics\u003c/summary\u003e\n    \u003ch3\u003eCIFAR-10\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eModel\u003c/th\u003e\n                \u003cth\u003eMetric\u003c/th\u003e\n                \u003cth colspan=\"5\"\u003eNoise Type/Ratio\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctr\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e20% sym\u003c/th\u003e\n                \u003cth\u003e50% sym\u003c/th\u003e\n                \u003cth\u003e80% sym\u003c/th\u003e\n                \u003cth\u003e90% sym\u003c/th\u003e\n                \u003cth\u003e40% asym\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctbody\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eSOTA\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e96.100%\u003c/td\u003e\n                    \u003ctd\u003e94.600%\u003c/td\u003e\n                    \u003ctd\u003e93.200%\u003c/td\u003e\n                    \u003ctd\u003e76.000%\u003c/td\u003e\n                    \u003ctd\u003e93.400%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e95.700%\u003c/td\u003e\n                    \u003ctd\u003e94.400%\u003c/td\u003e\n                    \u003ctd\u003e92.900%\u003c/td\u003e\n                    \u003ctd\u003e75.400%\u003c/td\u003e\n                    \u003ctd\u003e92.100%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003e\u003cstrong\u003eOurs\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e96.540%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e95.640%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e93.770%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e91.880%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e94.640%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e96.327%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e95.417%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e93.641%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e91.760%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e94.258%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n            \u003c/tbody\u003e\n        \u003c/thead\u003e\n    \u003c/table\u003e\n    \u003ch3\u003eCIFAR-100\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eModel\u003c/th\u003e\n                \u003cth\u003eMetric\u003c/th\u003e\n                \u003cth colspan=\"5\"\u003eNoise Type/Ratio\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctr\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e\u003c/th\u003e\n                \u003cth\u003e20% sym\u003c/th\u003e\n                \u003cth\u003e50% sym\u003c/th\u003e\n                \u003cth\u003e80% sym\u003c/th\u003e\n                \u003cth\u003e90% sym\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctbody\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003eSOTA\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e77.300%\u003c/td\u003e\n                    \u003ctd\u003e74.600%\u003c/td\u003e\n                    \u003ctd\u003e60.200%\u003c/td\u003e\n                    \u003ctd\u003e31.500%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e76.900%\u003c/td\u003e\n                    \u003ctd\u003e74.200%\u003c/td\u003e\n                    \u003ctd\u003e59.600%\u003c/td\u003e\n                    \u003ctd\u003e31.000%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd rowspan=\"2\"\u003e\u003cstrong\u003eOurs\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003eHighest\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e79.890%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e77.640%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e66.360%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e41.200%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eLast 10\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e79.511%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e77.522%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e66.046%\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e40.895%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n            \u003c/tbody\u003e\n        \u003c/thead\u003e\n    \u003c/table\u003e\n    \u003ch3\u003eClothing1M\u003c/h3\u003e\n    \u003ctable\u003e\n        \u003cthead\u003e\n            \u003ctr\u003e\n                \u003cth\u003eModel\u003c/th\u003e\n                \u003cth\u003eAccuracy\u003c/th\u003e\n            \u003c/tr\u003e\n            \u003ctbody\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003eSOTA\u003c/td\u003e\n                    \u003ctd\u003e74.760%\u003c/td\u003e\n                \u003c/tr\u003e\n                \u003ctr\u003e\n                    \u003ctd\u003e\u003cstrong\u003eOurs\u003c/strong\u003e\u003c/td\u003e\n                    \u003ctd\u003e\u003cstrong\u003e75.109%\u003c/strong\u003e\u003c/td\u003e\n                \u003c/tr\u003e\n            \u003c/tbody\u003e\n        \u003c/thead\u003e\n    \u003c/table\u003e\n\u003c/details\u003e\n\n## Training Locally\n\n\u003e The source code is heavily reliant on CUDA. Please make sure that you have the newest version of Pytorch and a compatible version of CUDA installed. Using older versions may exhibit inconsistent performance.\n\u003e\n\u003e [Download Pytorch](https://pytorch.org/get-started/locally/)\n\u003e /\n\u003e [Download CUDA](https://developer.nvidia.com/cuda-downloads)\n\u003e\n\u003e Other requirements are included in `requirements.txt`.\n\n\n\u003e **Reproducibility**\n\u003e \n\u003e At particularly high noise ratios (ex. 90% on CIFAR-10), results may vary across training runs.\n\u003e We are aware of this issue, and are exploring ways to yield more consistent results.\n\u003e We will publish any findings (consistently performant configurations, improved procedures, etc.) both in this repository and in continuations of this work.\n\nAll training configurations and parameters are controlled via the `presets.json` file. Configurations can contain infinite subconfigurations, and settings specified in subconfigurations always override the parent.\n\nTo train locally, first add your local machine to the `presets.json`:\n```javascript\n{\n    // ... inside the root scope\n    \"machines\": { // list of machines\n        \"localPC\": { // name for your local PC, can be anything\n            \"checkpoint_path\": \"./localPC_checkpoints\"\n        }\n    },\n    \"configs\": {\n        \"c10\": { // cifar-10 dataset\n            \"machines\": { // list of machines\n                \"localPC\": { // local PC name\n                    \"data_path\": \"/path/to/your/dataset\"\n                    // path to dataset (python) downloaded from:\n                    // https://www.cs.toronto.edu/~kriz/cifar.html\n                }\n                // ... keep all other machines unchanged\n            }\n            // ... keep all other config values unchanged\n        }\n        // ... keep all other configs unchanged\n    }\n    // ... keep all other global values unchanged\n}\n```\n\nA \"preset\" is a specific configuration branch. For example, if you would like to run `train_cifar.py` with the preset  `root -\u003e c100 -\u003e 90sym -\u003e AugDesc-WS` on your machine named `localPC`, you can run the following command:\n```bash\npython train_cifar.py --preset c100.90sym.AugDesc-WS --machine localPC\n```\nThe script will begin training the preset specified by the `--preset` argument. Progress will be saved in the appropriate directory in your specified `checkpoint_path`. Additionally, if the `--machine` flag is ommitted, the training script will look for the dataset in the `data_path` inherited from parent configurations.\n\nHere are some abbreviations used in our `presets.json`:\n| Abbreviation | Meaning                   |\n| :----------- | :------------------------ |\n| `c10`        | CIFAR-10                  |\n| `c100`       | CIFAR-100                 |\n| `c1m`        | Clothing1M                |\n| `sym`        | Symmetric Noise           |\n| `asym`       | Asymmetric Noise          |\n| `SAW`        | Strongly Augmented Warmup |\n| `WAW`        | Weakly Augmented Warmup   |\n| `RandAug`    | RandAugment               |\n\n\n## Citations\nPlease cite the following:\n```\n@InProceedings{Nishi_2021_CVPR,\n    author    = {Nishi, Kento and Ding, Yi and Rich, Alex and {H{\\\"o}llerer, Tobias},\n    title     = {Augmentation Strategies for Learning With Noisy Labels},\n    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},\n    month     = {June},\n    year      = {2021},\n    pages     = {8022-8031}\n}\n```\n\n## Extras\n\nExtra bits of unsanitized code for plotting, training, etc. can be found in the [Aug-for-LNL-Extras repository](https://github.com/KentoNishi/Aug-for-LNL-Extras).\n\n## Additional Info\nThis repository is a fork of [the official DivideMix implementation](https://github.com/LiJunnan1992/DivideMix).\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fkentonishi%2Faugmentation-for-lnl","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fkentonishi%2Faugmentation-for-lnl","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fkentonishi%2Faugmentation-for-lnl/lists"}